---
# @package _global_

defaults:
  - override /mapping: null
  - override /criterion: cross_entropy
  - override /callbacks: htr
  - override /datamodule: iam_extended_paragraphs
  - override /network: null
  - override /model: null
  - override /lr_schedulers: null
  - override /optimizers: null

epochs: &epochs 200
ignore_index: &ignore_index 3
num_classes: &num_classes 58
max_output_len: &max_output_len 682
summary: [[1, 1, 576, 640], [1, 682]]

criterion:
  ignore_index: *ignore_index
  # label_smoothing: 0.1

mapping: &mapping
  mapping:
    _target_: text_recognizer.data.mappings.emnist.EmnistMapping
    extra_symbols: ["\n"]

callbacks:
  stochastic_weight_averaging:
    _target_: pytorch_lightning.callbacks.StochasticWeightAveraging
    swa_epoch_start: 0.75
    swa_lrs: 1.0e-5
    annealing_epochs: 10
    annealing_strategy: cos
    device: null

optimizers:
  radam:
    _target_: torch.optim.RAdam
    lr: 1.5e-4
    betas: [0.9, 0.999]
    weight_decay: 0
    eps: 1.0e-8
    parameters: network

lr_schedulers:
  network:
    _target_: torch.optim.lr_scheduler.ReduceLROnPlateau
    mode: min
    factor: 0.5
    patience: 10
    threshold: 1.0e-4
    threshold_mode: rel
    cooldown: 0
    min_lr: 1.0e-6
    eps: 1.0e-8
    verbose: false
    interval: epoch
    monitor: val/loss

datamodule:
  _target_: text_recognizer.data.iam_extended_paragraphs.IAMExtendedParagraphs
  batch_size: 6
  num_workers: 12
  train_fraction: 0.8
  pin_memory: true
  << : *mapping

encoder: &encoder
  _target_: text_recognizer.networks.efficientnet.efficientnet.EfficientNet
  arch: b0
  stochastic_dropout_rate: 0.2
  bn_momentum: 0.99
  bn_eps: 1.0e-3
  depth: 7

rotary_embedding: &rotary_embedding
  rotary_embedding:
    _target_: >
      text_recognizer.networks.transformer.embeddings.rotary.RotaryEmbedding
    dim: 64

attn: &attn
  dim: &hidden_dim 512
  num_heads: 4
  dim_head: 64
  dropout_rate: &dropout_rate 0.4

decoder: &decoder
  _target_: text_recognizer.networks.transformer.decoder.Decoder
  depth: 6
  has_pos_emb: true
  block:
    _target_: text_recognizer.networks.transformer.decoder.DecoderBlock
    self_attn:
      _target_: text_recognizer.networks.transformer.attention.Attention
      << : *attn
      causal: true
      << : *rotary_embedding
    cross_attn:
      _target_: text_recognizer.networks.transformer.attention.Attention
      << : *attn
      causal: false
    norm:
      _target_: text_recognizer.networks.transformer.norm.RMSNorm
      dim: *hidden_dim
    ff:
      _target_: text_recognizer.networks.transformer.mlp.FeedForward
      dim: *hidden_dim
      dim_out: null
      expansion_factor: 2
      glu: true
      dropout_rate: *dropout_rate

pixel_pos_embedding: &pixel_pos_embedding
  _target_: >
    text_recognizer.networks.transformer.embeddings.axial.AxialPositionalEmbedding
  dim: *hidden_dim
  shape: &shape [18, 20]

network:
  _target_: text_recognizer.networks.conv_transformer.ConvTransformer
  input_dims: [1, 1, 576, 640]
  hidden_dim: *hidden_dim
  num_classes: *num_classes
  pad_index: *ignore_index
  encoder:
    << : *encoder
  decoder:
    << : *decoder
  pixel_pos_embedding:
    << : *pixel_pos_embedding

model:
  _target_: text_recognizer.models.transformer.TransformerLitModel
  << : *mapping
  max_output_len: *max_output_len
  start_token: <s>
  end_token: <e>
  pad_token: <p>

trainer:
  _target_: pytorch_lightning.Trainer
  stochastic_weight_avg: true
  auto_scale_batch_size: binsearch
  auto_lr_find: false
  gradient_clip_val: 0.5
  fast_dev_run: false
  gpus: 1
  precision: 16
  max_epochs: *epochs
  terminate_on_nan: true
  weights_summary: null
  limit_train_batches: 1.0
  limit_val_batches: 1.0
  limit_test_batches: 1.0
  resume_from_checkpoint: null
  accumulate_grad_batches: 2
  overfit_batches: 0